why there is a complex sync between /var/lib/nfs/etab in user-mode and export_table in kernel mode?

why there is a complex sync between /var/lib/nfs/etab in user-mode and export_table in kernel mode?

[lioupayphone]

Hello, linux-nfs

i found that an entry in /var/lib/nfs/etab (etab for abbr)  was written to exprot_table of kernel-mode via a proc-file when a client requests mnt this entry.                                       
when the entry in export_table of kernel-mode was expired (eg : reached its expiration time), an upcall machanism is able to make sure that the newly entry from etab of user-mode was written to export_table of kernel-mode. ie, the entry in export_table of kernel-mode was updated by the one of user-mode.

if the content listed above is true, i think linux nfs is too complex. and why not remove the etab from user-mode?
in a word, i think it is easy to just maintain export_table of kernel-mode. we can directly scan /etc/exports and make up a complete export-entry, then write it to the export_table of kernel-mode. the entries in export_table of kernel-mode would never be expired unless we explicitly remove it from the export_table of kernel-mode (a proc MUST be provided for meeting this requirement).

when the server is rebooted and nfsd proc-filesystem is ready, just re-do "scanning the /etc/exports and make up complete export-entries, then write them into export_table of kernel-mode".  

i aims to 1) remove the complex upcall machanism ,  2) get rid of the couple of /var/lib/nfs/etab in user-mode and export_table in kernel-mode.

anyone can give me some suggestions? or points out what's wrong i am.
thank you!
 				
Best Regards! 
lioupayphone

Re: why there is a complex sync between /var/lib/nfs/etab in user-mode and export_table in kernel mode?

[Steve Dickson]

lioupayphone wrote:
> Hello, linux-nfs
> 
> i found that an entry in /var/lib/nfs/etab (etab for abbr)  was written to exprot_table of kernel-mode via a proc-file when a client requests mnt this entry.                                       
> when the entry in export_table of kernel-mode was expired (eg : reached its expiration time), an upcall machanism is able to make sure that the newly entry from etab of user-mode was written to export_table of kernel-mode. ie, the entry in export_table of kernel-mode was updated by the one of user-mode.
> 
> if the content listed above is true, i think linux nfs is too complex. and why not remove the etab from user-mode?
> in a word, i think it is easy to just maintain export_table of kernel-mode. we can directly scan /etc/exports and make up a complete export-entry, then write it to the export_table of kernel-mode. the entries in export_table of kernel-mode would never be expired unless we explicitly remove it from the export_table of kernel-mode (a proc MUST be provided for meeting this requirement).
> 
> when the server is rebooted and nfsd proc-filesystem is ready, just re-do "scanning the /etc/exports and make up complete export-entries, then write them into export_table of kernel-mode".  
> 
> i aims to 1) remove the complex upcall machanism ,  2) get rid of the couple of /var/lib/nfs/etab in user-mode and export_table in kernel-mode.
> 
> anyone can give me some suggestions? or points out what's wrong i am.
The /var/lib/nfs/etab file is the way the exportfs command communicates 
with the mountd daemon. 'exportfs -ar' causes the exports in /etc/exports
to be parsed and written to the etab. When a mount request is received
mountd reads the etab to build its internal export table. Ultimately
when the mount is successful, the kernel writes the mount to
/proc/net/rpc/nfsd.export/content.

So that's how the etab fits in the grand scheme of things... 

I believe what you are suggesting is simply pumping down
all the exports in /etc/exports directly to the kernel.
This ideas has been discussed and I believe the conclusion
we came to was; why pump down thousands of exports that
may or may not used. Dynamically building the kernel export
data just seem to make more sense in our case...

I hope this helps...

steved.

Re: Re: why there is a complex sync between /var/lib/nfs/etab in user-modeand export_table in kernel mode?

[lioupayphone]

Hello, Steve Dickson.
you wrote at 23:54:00 on 2008-12-18 :
>"Re: why there is a complex sync between /var/lib/nfs/etab in user-modeand  export_table in kernel mode?"
>lioupayphone wrote:
>> Hello, linux-nfs
>> 
>> i found that an entry in /var/lib/nfs/etab (etab for abbr)  was written to exprot_table of kernel-mode via a proc-file when a client requests mnt this entry.                                       
>> when the entry in export_table of kernel-mode was expired (eg : reached its expiration time), an upcall machanism is able to make sure that the newly entry from etab of user-mode was written to export_table of kernel-mode. ie, the entry in export_table of kernel-mode was updated by the one of user-mode.
>> 
>> if the content listed above is true, i think linux nfs is too complex. and why not remove the etab from user-mode?
>> in a word, i think it is easy to just maintain export_table of kernel-mode. we can directly scan /etc/exports and make up a complete export-entry, then write it to the export_table of kernel-mode. the entries in export_table of kernel-mode would never be expired unless we explicitly remove it from the export_table of kernel-mode (a proc MUST be provided for meeting this requirement).
>> 
>> when the server is rebooted and nfsd proc-filesystem is ready, just re-do "scanning the /etc/exports and make up complete export-entries, then write them into export_table of kernel-mode".  
>> 
>> i aims to 1) remove the complex upcall machanism ,  2) get rid of the couple of /var/lib/nfs/etab in user-mode and export_table in kernel-mode.
>> 
>> anyone can give me some suggestions? or points out what's wrong i am.
>The /var/lib/nfs/etab file is the way the exportfs command communicates 
>with the mountd daemon. 'exportfs -ar' causes the exports in /etc/exports
>to be parsed and written to the etab. When a mount request is received
>mountd reads the etab to build its internal export table. Ultimately
>when the mount is successful, the kernel writes the mount to
>/proc/net/rpc/nfsd.export/content.
>
>So that's how the etab fits in the grand scheme of things... 
>
>I believe what you are suggesting is simply pumping down
>all the exports in /etc/exports directly to the kernel.
>This ideas has been discussed and I believe the conclusion
>we came to was; why pump down thousands of exports that
>may or may not used. Dynamically building the kernel export
>data just seem to make more sense in our case...
>
>I hope this helps...
:-) thank you very much. i think your comments are very clear and very helpful for me.
but i still think rpc.mountd is somewhat complex. we all know that a daemon in user-mode is likely to be killed. rpc.mountd is no exception. once rpc.mountd was killed, there are no chance for the export_table to be updated via upcall.

It cannot be denied that putting the etab into kernel mode is wastful for memory. but i still think it is an easy  and stable method.

thanks again. 
:-)

>
>steved.
>

Best Regards!
lioupayphone

Re: Re: why there is a complex sync between /var/lib/nfs/etab in user-modeand export_table in kernel mode?

[J. Bruce Fields]

On Fri, Dec 19, 2008 at 10:50:27PM +0800, lioupayphone wrote:
> >I hope this helps...
> :-) thank you very much. i think your comments are very clear and very
> helpful for me.  but i still think rpc.mountd is somewhat complex. we
> all know that a daemon in user-mode is likely to be killed. rpc.mountd
> is no exception. once rpc.mountd was killed, there are no chance for
> the export_table to be updated via upcall.
> 
> It cannot be denied that putting the etab into kernel mode is wastful
> for memory. but i still think it is an easy  and stable method.

Even with the export table in memory, rpc.mountd will still be needed
for ip-address lookup (e.g. if a filesystem is exported to *.umich.edu,
and the kernel gets an rpc from 141.211.133.1, then it needs to do an
upcall to determine whether the given ip address matches that wildcard).

However it's true that there might be advantages to having the export
table in-kernel.

--b.